Troubleshooting

Errors are grouped by the message Sira prints.

Playwright cannot find a browser

❌ Failed to scrape job posting from URL: … Executable doesn't exist at …
💡 Tip: Ensure the URL is publicly accessible and contains a valid job posting.

The Playwright Python package is installed, but the actual browser binary is a separate download. Install it once:

uv run sira setup        # or just `sira setup` when installed from PyPI

On Linux you may also need the shared libraries the browser links against:

uv run playwright install --with-deps chromium

❌ Failed to scrape job posting from URL

The scraper drives a real headless browser, so it can read JavaScript-rendered postings. It still fails when:

Cause	What to do
The posting is behind a login	Save the posting text to a file and open an issue — Sira currently only scrapes public URLs.
The page is bot-protected	Some boards block headless browsers outright. Try the employer's own careers page rather than the aggregator.
The page is slow	The default timeout is 30 seconds per navigation. A very slow page can exceed it; retry.
The URL redirects to a search page	The posting has probably expired.

⚠️ Potential prompt-injection content detected in the job posting

The page contains text aimed at an AI reader — an instruction override, a hidden block, a request to rate the candidate highly — or a phrase that exists in the HTML but not in the visible text. The run continues; the warning lists only category names (see the agent reference for the full table). Read the tailored CV and report with extra care, and compare the requirements Sira extracted against the posting as a human sees it. False positives happen on security-engineering postings that discuss these attacks; the warning is advisory, nothing is blocked.

⚠️ Prompt-injection classifier unavailable (…)

You installed the guard extra but the local model could not be loaded — usually a missing Hugging Face token for the gated default model, no network on first run, or not enough memory. Sira falls back to the regex scan. Fix the cause, or set SIRA_GUARD_MODEL=protectai/deberta-v3-base-prompt-injection-v2 for an ungated model, or set SIRA_GUARD_CONSENT=no to silence the layer.

❌ Job posting scraped but content is empty

The page loaded, but nothing survived extraction. This normally means the posting body is rendered behind an interaction (an "expand" button) or the page is a listing rather than a single posting. Use the URL of the posting itself.

❌ Error: Job URL must start with http:// or https://

The first argument is the URL, the second is the resume path — in that order. A local file path in the first position produces this message.

❌ Resume file not found at …

The path does not exist. ~ is expanded, but shell quoting still matters:

uv run sira tailor <JOB_URL> "~/Documents/my resume.pdf"

❌ Failed to convert resume: …

Only .md, .docx, and .pdf are accepted. Some PDFs — scans, or heavily designed layouts — extract badly or not at all. Two things to try:

1. Export the resume to .docx from the original editor and pass that instead.
2. Run with --debug and read resume_converted.md in the output directory. That file is exactly what the parser saw. If it is empty or scrambled, the problem is conversion, not the model.

❌ Resume content is empty

The file exists but converted to nothing. Almost always an image-only PDF, which has no extractable text layer. Convert it to text first, or use the original document.

❌ Original resume not found at recorded path: …

re-tailor looks the original resume up at the path recorded when you first ran tailor. If you moved or renamed that file, point at the new location:

uv run sira re-tailor <JOB_ID> "…" --resume-path ~/Documents/resume.pdf

❌ Job not found: <id>

The job ID is the UUID printed at the end of a successful tailor run, and it is looked up in the memory database. Since 1.5 that database is shared across working directories; before 1.5 it was memory/resume_memory.sqlite3 relative to where you ran tailor, and Sira moves such a file into the data directory the first time it runs from that place. If the ID came from a run in another directory on an older release, run sira once from that directory so its database is moved.

❌ No job posting content stored for this job

The job row predates the column that stores the posting text, so re-tailor has nothing to work from. Re-run tailor on the original URL.

Authentication errors from the model provider

AuthenticationError: Incorrect API key provided …

Each provider reads its own environment variable — see Models and providers. Two cases catch people out:

• No key at all. Agents are built at import time with a placeholder key so that importing Sira never fails. The error only appears at the first real call.
• --fast with a non-OpenAI model. The fast tier stays on openai:gpt-5-nano, so an OPENAI_API_KEY is still required. Drop --fast to use a single provider.

Ollama: connection refused, or a missing base URL

PydanticAI has no default endpoint for Ollama. Export it explicitly:

export OLLAMA_BASE_URL=http://localhost:11434/v1

Then confirm the daemon is up (ollama list) and the model is pulled (ollama pull llama3).

The audit keeps failing

The auditor rejects a draft that invents facts, leans on clichés, or drops hyperlinks. Give the loop more room:

uv run sira tailor <JOB_URL> <RESUME_PATH> --write-attempts 3 --review-iterations 2

If it still fails, read the report's Audit Summary — the reason is usually specific, and re-tailor with an instruction addressing it works better than more retries.

The terminal output looks garbled

The live dashboard and some plain print() calls in the workflow can interleave in an interactive terminal. The run is unaffected; only the drawing is. Use --verbose for clean streaming output, or pipe the output — in a non-TTY the dashboard degrades to line-by-line logging.

The run is slow or expensive

Start with --fast, then tune:

uv run sira tailor <JOB_URL> <RESUME_PATH> \
  --no-quality-gate --write-attempts 1 --review-iterations 0

The trade-offs behind each flag are in Controlling cost.

Nothing here matches

Run with --verbose to see each agent's input and output as it streams, and open an issue at Tiqni/sira/issues with that output.